Jump to content

  • Log In with Google      Sign In   
  • Create Account

What graphics quality in real-time (~30 fps) we can get on 1 PFlops gpgpu with 16-20 Gb memory ?


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
9 replies to this topic

#1 xma   Members   -  Reputation: 210

Like
1Likes
Like

Posted 16 December 2012 - 06:31 AM

What graphics quality in real-time (~30 fps) we can get on 1 PFlops gpgpu with 16-20 Gb memory ?


what system needs for real-time for scenes below (multiCISC, RISK? PFlops ?) ?

(where Sarah Kerrigan),





Sponsor:

#2 MarkS   Prime Members   -  Reputation: 882

Like
2Likes
Like

Posted 16 December 2012 - 01:32 PM

I don't think this is the correct forum for speculation on non-existent hardware, but I'll bite...

First off, any graphics card with that much memory is going to have major issues. My motherboard only supports 16GB of RAM and I have 16GB of RAM installed. I also have a graphics card with 1GB of RAM. That means that I have 15GB available to system use and 1GB for graphics. If I were to install a graphics card with 16GB of RAM, I would have no RAM for the OS or applications. Keep in mind that the memory controller on the newer intel processors has a 48 GB limit. That is total system RAM. Server processors and motherboards can address more, but memory size is a minor concern.

As to the PetaFLOPS, that is meaningless. In order for a GPU or CPU to operate on floating point instructions at that speed, it would have to be able move the instructions across the bus at no less than that speed. We lack the technology to do this. The world's fastest super computer operates at those speeds, but it is constructed of thousands of highly specialized motherboards containing off-the-shelf CPUs and GPUs, all of which operate at much lower speeds and all in a connected massively parallel system. The petaFLOP speed of a super computer is due to the massively parallel configuration and not because any one component operates even close to that speed.

To be blunt, such speculation is so far out there that it simply lacks merit. We simply do not have the technology to create a GPU with these capabilities, and even if we did, it would be so bottlenecked by current CPU and memory technologies that it wouldn't operate any faster than current GPUs. This is akin to asking what kind of graphics quality could we expect if we had access to the starship Enterprise's main computer.

#3 xma   Members   -  Reputation: 210

Like
-1Likes
Like

Posted 16 December 2012 - 02:01 PM

MarkS,
2016 - x24 gpgpu (35 TFlops) ~0,84 Pflops
Looks good ! ))

for this trailers for real time - can up to 100 PFlops . But I think that is not enough.

So, what kind of graphics we can get on such systems ?

#4 MarkS   Prime Members   -  Reputation: 882

Like
0Likes
Like

Posted 16 December 2012 - 02:13 PM

I really think you lack any understanding of the hardware behind the software. GPGPU is not some sort of magic bullet. It can only process data as fast as it can pull it across its bus and no bus technology can operate at those speeds. Moreover, it would be terribly bottlenecked as the CPU could not supply the GPU with data fast enough. The technology doesn't exist. Or, to put that another way, it is so far beyond current tech that we cannot imagine the technology required to accomplish it.

I'm not sure what you mean by "x24 gpgpu (35 TFlops) ~0,84 Pflops", but you mention 2016. That is time frame highly speculative and without substance.

To answer the other question, "So, what kind of graphics we can get on such systems ?", I simply do not know and I doubt there is a person in the industry that could answer that. That would require speculation, not only on the speed of the GPU, but the underlying technology (CPU, RAM, bus, etc.) required to support it. First, we'd have to conjure up a system based on non-existent technology before we could begin to answer the question.

#5 xma   Members   -  Reputation: 210

Like
0Likes
Like

Posted 16 December 2012 - 02:19 PM

MarkS, "x24 gpgpu (35 TFlops) ~0,84 Pflops" - 3 servers with 8 top videocards in each in 2016 is 0,84 Pflops .
Can we render something interesting on it (more interesting than on one videocard) ? samaritian demo render on 3 gtx580 for example . and it looks good .

Edited by xma, 16 December 2012 - 02:20 PM.


#6 MarkS   Prime Members   -  Reputation: 882

Like
0Likes
Like

Posted 16 December 2012 - 03:54 PM

MarkS, "x24 gpgpu (35 TFlops) ~0,84 Pflops" - 3 servers with 8 top videocards in each in 2016 is 0,84 Pflops ..


This isn't something that you have to wait for. Massive render farms exist with current technology that far exceed what you posted. Pixar and Weta Workshop have massive render farms with hundreds, if not thousands, of servers containing several GPUs each, all arranged in parallel. Even with this massive amount of computing, all of which is geared towards graphics, it still takes hours per frame to render.

In this case, it comes down to a software issue. Most programmers do not know how to write truly parallel code and GPU drivers do a fairly poor job at it. Because of this, the load is not balanced across the GPUs. 8 GPUs does not equal 8x performance. And there are still bandwidth issues at play.

You are falling for a common computing misconception. People think that throwing more processors and/or faster processors at a task will speed up the task exponentially. The truth is that it wouldn't even speed up the task linearly. It is similar to throwing money at world hunger, thinking that the more money you use the less starving people there will be. That is only true is all of the money gets to the starving people (it wont and never does) and if money is the sole reason for their lack of food (most often not). Point being, it misses the small and yet incredibly important technical details that hinder the plan from working.

Edited by MarkS, 16 December 2012 - 04:08 PM.


#7 GeniusPooh   Members   -  Reputation: 157

Like
0Likes
Like

Posted 16 December 2012 - 10:57 PM

It's too hard to estimate :)

and there is so many possiblity

but.. you can calcuate by radiosity and raytracing by that math ability..

but mostly we earn much more than those non realtime method..

Beauty is only skin deep , ugly goes to bones

 

World's only 3D engine tunner and 3D engine guru.

 

and real genius inventor :) but very kind warm heart .. and having serious depression for suffering in Korea

 

www.polygonart.co.kr ( currently out dated and only Korean will change to English and new stuff when I get better condition :) sorry for that)


#8 codyrigney92   Members   -  Reputation: 103

Like
0Likes
Like

Posted 17 December 2012 - 01:04 AM

I don't know why everyone is bringing this down.

First of all, the hardware you specify is theoretically possible, but as of today no manufacturer has done it.(For very good reasons).

Second of all, those videos you posted can be rendered in real time, but it all depends on the details. Resolution? Different effects such as Depth of Field, etc? Those videos were rendered through Ray Traced graphics and with some powerful at home hardware, you could render a pretty good looking scene at a decent FPS. Not as good looking as the videos you posted.

Some real time ray tracing:



The pixar scenes that one user was talking about are rendered in a high resolution, probably 4K. However they don't take several hours per frame otherwise it would take over 100 years. You can render a complex high res scene on a modern computer in ~30 secs or less.

Unless they can get rasterized graphics to simulate light as good as ray tracing, ray tracing will eventually be implemented inside GPUs in the future. They have GPUs with 3,000 cores. If each core could perform calculations for ray tracing, you would have real time. Current GPU hardware limitations don't make that part so easy.

#9 MJP   Moderators   -  Reputation: 11385

Like
0Likes
Like

Posted 17 December 2012 - 01:52 AM

However they don't take several hours per frame otherwise it would take over 100 years.


It absolutely can take hours to render a single frame, and it doesn't take "100 years" to render a movie because they render lots of frames simultaneously. It's not like Pixar just has one PC in closet that does all of their rendering. Like the user above mentioned they have a "farm" of computers that splits up rendering jobs as they come in.

Unless they can get rasterized graphics to simulate light as good as ray tracing, ray tracing will eventually be implemented inside GPUs in the future. They have GPUs with 3,000 cores. If each core could perform calculations for ray tracing, you would have real time. Current GPU hardware limitations don't make that part so easy.


You seem to be rather mixed up about what ray tracing and rasterization really accomplish, as well as what's possible on current GPU's (here's a hint: people have been doing ray tracing on GPU's for years now).

Edited by MJP, 17 December 2012 - 01:56 AM.


#10 xma   Members   -  Reputation: 210

Like
0Likes
Like

Posted 17 December 2012 - 06:08 AM

samaritian demo looks good than radiosity .. but some fakes are visible .


and only 5-10 TFlops for real-time needed.




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS